[Issue]: Vectors collapse (non-discriminative) with the xclbins shipped for ≥0.9.45 #789

Problem Description

Summary
With FastFlowLM ≥0.9.45, embed-gemma:300m (--embed 1) returns deterministic but non-discriminative embeddings. Unrelated texts score about as high as similar ones, and the vectors are uncorrelated with
those from 0.9.42. Copying the four Embedding-Gemma-300M-NPU2 xclbins from 0.9.42 into a 1.0.8 install gives correct embeddings again. This points at the refreshed xclbins, not the host library or the
open-source embedding code. It looks like the same problem as #732 and #661.

Environment
- AMD Ryzen AI Max+ 395 (Strix Halo), NPU firmware 1.1.2.65, amdxdna 0.16, Linux 7.2.10 (Fedora 43)
- FastFlowLM from the Fedora COPR (alessandrolattao/fastflowlm). Kernels and libs come from flm-fetch-kernels, which fetches the matching release's kernel files.
- Model embed-gemma:300m, served with flm serve qwen3-it:4b --embed 1, queried through /v1/embeddings

Method
- Discrimination test: I embedded six texts and compared cosine similarity: two similar pairs (FreeIPA password reset, and Ceph/CloudStack storage full) and unrelated texts (a cake recipe, a Wazuh agent message).
- Fidelity test: I re-embedded 60 texts that already had vectors stored from an earlier run on 0.9.42 and compared them.

Results

│ FastFlowLM │ Similar pairs │ Unrelated pairs │ Cosine vs stored 0.9.42 vectors (60 texts) │
│ 0.9.42-5 │ 0.86 / 0.52 │ 0.09 / 0.08 / 0.18 │ mean 0.9998 (min 0.9996) │
│ 0.9.43-2 │ 0.86 / 0.52 │ 0.09 / 0.08 / 0.18 │ mean 0.9998 (min 0.9996) │
│ 0.9.45-2 │ 0.74 / 0.83 │ 0.79 / 0.75 / 0.72 │ mean −0.035 (min −0.129) │
│ 1.0.8-1 │ 0.75 / 0.83 │ 0.79 / 0.75 / 0.73 │ mean −0.033 (min −0.129) │

Things I ruled out
- Model weights: I re-pulled embed-gemma:300m, and the files are byte-identical (md5) to the July copy.
- Prompt format: neither a query prefix nor a document prefix fixed it, and passing dimensions: 768 made no difference. The same text embedded twice gave cosine 1.0000, so the output is deterministic, not random.
- Batching: one text at a time and a batch of six gave identical results.
- Stale libraries: I cleared /opt/fastflowlm/lib* and share/flm/xclbins and re-ran flm-fetch-kernels before each test.

Operating System

Fedora 43

CPU

AMD RYZEN AI MAX+ 395 w/ Radeon 8060S

GPU

AMD Radeon 8060S Graphics

ROCm Version

ROCm 10

Installation Method

Other (describe in Additional Information)

Installed ROCm Packages / Versions

Paste output here

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of rocminfo --support

Paste output here

Additional Information

No response